Harvey’s Margins, Are Insects Worth More Than Humans?, Meta Tests Human Help for Muse | Diet TBPN

Harvey’s sudden margin deterioration captures a new problem for AI application companies: usage can become far more valuable while unit economics get worse. Reasoning models and agentic workflows drove a twentyfold increase in token consumption, briefly turning gross margin from roughly 50% to about negative 50%.

Rebuilding the inference stack

Harvey says it returned to positive gross margin within one quarter. The response combined model routing, harness improvements, spend controls and post-training an open-weight model. The episode also exposes the mismatch between seat-based pricing and actual consumption: the same user instruction can now trigger broad retrieval, multiple reasoning steps and many model calls.

The moral arithmetic of insects

The hosts then examine an essay claiming insects may matter more than people in aggregate. Its case depends on scale rather than equal individual experience: if insects feel even a tiny fraction of human pain, their vast numbers could produce more total suffering. The practical implications remain unclear, especially compared with interventions tied to farming or other preventable harms.

Humans remain inside Meta’s agent

Meta is testing human contractors to handle some phone calls that Muse cannot yet complete reliably. This looks less like a permanent service model than a learning loop: humans cover difficult cases while producing demonstrations that can improve future models.

The experiment points to a larger commerce battle. Amazon controls logistics and a large advertising business, Shopify can use agent integrations to strengthen Shop Pay, and consumer assistants want to own the purchasing interface. Marketplaces must decide whether to open their APIs or defend conversion, basket size and ad revenue.

What to watch

• Whether Harvey can sustain software-like margins as agent usage grows.
• Dynamic routing between frontier and specialized open-weight models.
• Human-in-the-loop data collection for voice and transaction agents.
• Access terms imposed by Amazon, Shopify and other commerce platforms.

Source

• Chaîne: TBPN
• Vidéo source: https://www.youtube.com/watch?v=-96mtSZkEL0